Introduction

Xgboost, eXtreme Gradient Boosting，是目前在 Kaggle 數據分析平台上奪冠常勝的演算法之一。

Xgboost Github

Basic Hands-on

這篇教學使用 Agaricus 資料集進行示範 (UCI Machine Learning Repository)，預測該香菇是否可食用

1. Input Type

  • Dense Matrix : matrix
  • Sparse Matrix : dgCMatrix
    （稀疏矩陣，雖有大量的 0 空值，但不會佔用記憶體，xgboost 有特別針對稀疏進行優化）
  • xgboost Exclusive : xgb.DMatrix （建議使用）

矩陣內僅接受 numeric 資料格式，不支援 categorical，若需進行 One-hot Encoding 轉成 dummy variable 處理

require(xgboost)

data(agaricus.train, package='xgboost')
data(agaricus.test, package='xgboost')

train <- agaricus.train
test <- agaricus.test

2. Observe Data Structure

訓練資料為 list 結構，含有資料 data 以及標籤 label

str(train)
## List of 2
##  $ data :Formal class 'dgCMatrix' [package "Matrix"] with 6 slots
##   .. ..@ i       : int [1:143286] 2 6 8 11 18 20 21 24 28 32 ...
##   .. ..@ p       : int [1:127] 0 369 372 3306 5845 6489 6513 8380 8384 10991 ...
##   .. ..@ Dim     : int [1:2] 6513 126
##   .. ..@ Dimnames:List of 2
##   .. .. ..$ : NULL
##   .. .. ..$ : chr [1:126] "cap-shape=bell" "cap-shape=conical" "cap-shape=convex" "cap-shape=flat" ...
##   .. ..@ x       : num [1:143286] 1 1 1 1 1 1 1 1 1 1 ...
##   .. ..@ factors : list()
##  $ label: num [1:6513] 1 0 0 1 0 0 0 1 0 0 ...


資料 data 為稀疏矩陣

class(train$data)[1]
## [1] "dgCMatrix"


標籤 label 代表訓練資料的結果，而我們透過 xgboost 預測 label

class(train$label)
## [1] "numeric"

3. Model Building

由於我們希望找出香菇為可食或不可食用，因此進行 objective = "binary:logistic" 二元分類

xgboost 詳細的參數設定，在進階篇章會深入探討

採用 dgCMatrix 丟入模型中

bstSparse <- xgboost(data = train$data, label = train$label, 
                     max_depth = 2, eta = 1, nthread = 2, nrounds = 2, objective = "binary:logistic")
## [1]  train-error:0.046522 
## [2]  train-error:0.022263


亦可轉換為 xgb.DMatrix

dtrain <- xgb.DMatrix(data = train$data, label = train$label)
bstDMatrix <- xgboost(data = dtrain, 
                      max_depth = 2, eta = 1, nthread = 2, nrounds = 2, objective = "binary:logistic")
## [1]  train-error:0.046522 
## [2]  train-error:0.022263

到這裡已經建立好基本的 xgboost 模型了

4. Prediction

將先前準備好的 test 測試資料集，投入到建立好的模型中進行預測


輸出結果為介在 [0, 1] 的機率，需再經由轉換得到最後的結果

pred <- predict(bstDMatrix, test$data)
head(pred)
## [1] 0.28583017 0.92392391 0.28583017 0.28583017 0.05169873 0.92392391


這邊我們以 0.5 做為分割點，以上代表 1, 反之為 0，得到模型預測的結果

prediction <- as.numeric(pred > 0.5)
head(prediction)
## [1] 0 1 0 0 0 1

5. Accuracy

建立 Contingency Table 觀察預測結果

table(prediction, test$label)
##           
## prediction   0   1
##          0 813  13
##          1  22 763


最後預測結果，準確度高達 98%

err <- mean(prediction != test$label)
print(paste("test-acc =", round(1- err, digits = 2)))
## [1] "test-acc = 0.98"


Advanced Techiniques

Data Preparation

在作者的文件當中，建議使用 xgb.DMatrix 資料型態，除了能夠讓 xgboost 更快的運算，同時也便於後續進行進階操作


xgb.DMatrix 不同於傳統資料格式，資料和標籤需分開儲存

dtrain <- xgb.DMatrix(data = train$data, label=train$label)
dtest <- xgb.DMatrix(data = test$data, label=test$label)


Information Extraction

label <- getinfo(dtest, "label")


Save / Load

xgb.DMatrix.save(dtrain, "dtrain.buffer")
dtrain2 <- xgb.DMatrix("dtrain.buffer")


Model Training

xgboost 和 xgb.train 皆是訓練模型函式

  • xgboost 為簡單易用版，能夠快速建立基本的模型
  • xgb.train 則提供更多參數調整，且只接受 xgbDMatrix


Parameter Setup

  • watchlist
watchlist <- list(train=dtrain, test=dtest)

bst <- xgb.train(data=dtrain, max_depth=2, eta=1, nthread = 2, nrounds=2, 
                 watchlist=watchlist, objective = "binary:logistic")
## [1]  train-error:0.046522    test-error:0.042831 
## [2]  train-error:0.022263    test-error:0.021726
# Multi Evaluation Metrics
bst <- xgb.train(data=dtrain, max_depth=2, eta=1, nthread = 2, nrounds=2, 
                 watchlist=watchlist, eval_metric = "error", eval_metric = "logloss", objective = "binary:logistic")
## [1]  train-error:0.046522    train-logloss:0.233376  test-error:0.042831 test-logloss:0.226686 
## [2]  train-error:0.022263    train-logloss:0.136658  test-error:0.021726 test-logloss:0.137874

train-error 和 test-error 輸出結果理論上需要相近，若是結果有明顯差異，需要重新檢視分割訓練和測試資料的狀況。

  • slice


Setup Self Evaluation Function

除了套件內提供的衡量標準，也可以視需求撰寫自己需要的標準

logregobj <- function(preds, dtrain) { 
  labels <- getinfo(dtrain, "label") 
  preds <- 1/(1 + exp(-preds))
  grad <- preds - labels
  hess <- preds * (1 - preds)

  return(list(grad = grad, hess = hess))
}

evalerror <- function(preds, dtrain) { 
  labels <- getinfo(dtrain, "label") 
  err <- sqrt(mean((preds-labels)^2))
  
  return(list(metric = "MSE", value = err)) 
}

dtest <- xgb.DMatrix(test$data, label = test$label)
watchlist <- list(eval = dtest, train = dtrain)
param <- list(max_depth = 2, eta = 1, silent = 1)

bst <- xgb.train(param, dtrain, nrounds = 2, watchlist, logregobj, evalerror, maximize = FALSE)
## [1]  eval-MSE:1.592293   train-MSE:1.595967 
## [2]  eval-MSE:2.405194   train-MSE:2.409772


View the Trees

xgb.dump(bst, with_stats = T)
##  [1] "booster[0]"                                                           
##  [2] "0:[f28<-9.53674e-07] yes=1,no=2,missing=1,gain=4254.64,cover=1530.58" 
##  [3] "1:[f55<-9.53674e-07] yes=3,no=4,missing=3,gain=1230.15,cover=869.044" 
##  [4] "3:leaf=1.30088,cover=763.292"                                         
##  [5] "4:leaf=-2.32415,cover=105.752"                                        
##  [6] "2:[f108<-9.53674e-07] yes=5,no=6,missing=5,gain=212.286,cover=661.535"
##  [7] "5:leaf=-2.58465,cover=649.08"                                         
##  [8] "6:leaf=1.48713,cover=12.4552"                                         
##  [9] "booster[1]"                                                           
## [10] "0:[f59<-9.53674e-07] yes=1,no=2,missing=1,gain=911.658,cover=726.268" 
## [11] "1:[f28<-9.53674e-07] yes=3,no=4,missing=3,gain=493.603,cover=707.047" 
## [12] "3:leaf=0.77417,cover=429.742"                                         
## [13] "4:leaf=-0.934515,cover=277.305"                                       
## [14] "2:leaf=-6.70608,cover=19.2211"
xgb.plot.tree(model = bst)


Feature Importance

查看資料集內各變數對預測結果的重要性程度

importance_matrix <- xgb.importance(model = bstDMatrix)
print(importance_matrix)
##                    Feature       Gain     Cover Frequency
## 1:               odor=none 0.67615484 0.4978746       0.4
## 2:         stalk-root=club 0.17135352 0.1920543       0.2
## 3:       stalk-root=rooted 0.12317241 0.1638750       0.2
## 4: spore-print-color=green 0.02931922 0.1461960       0.2
  • Gain : Representthe improvement in accuracy brought by a feature to the branches it is on.
  • Cover : Measures the relative quantity of observations concerned by a feature.
  • Frequency : A simpler way to measure the Gain. It just counts the number of times a feature is used in all generated trees.


xgb.plot.importance(importance_matrix = importance_matrix)

Linear Boosting

上面基本練習採用基於樹結構的方法，xgboost 也提供線性方法


線性方法常用在簡單的資料集，有效的抓出資料集內線性關係
但通常樹的方法普遍優於線性，找出變數間非線性的關係，建議嘗試兩種方式進行觀察

  • add booster : "gbLinear"
  • removing eta parameter
bst <- xgb.train(data=dtrain, booster = "gblinear", 
                 max_depth=2, nthread = 2, nrounds=2, 
                 watchlist=watchlist, 
                 eval_metric = "error", eval_metric = "logloss", objective = "binary:logistic")
## [1]  eval-error:0.005587 eval-logloss:0.162564   train-error:0.006909    train-logloss:0.162310 
## [2]  eval-error:0.000000 eval-logloss:0.056521   train-error:0.001842    train-logloss:0.059108


Model Saving / Loading

xgb.save(bst, "xgboost.model")
## [1] TRUE
bst_laod <- xgb.load("xgboost.model")
pred2 <- predict(bst_laod, test$data)


Reference